Re: Another RFC: regex in libiberty

["Zack Weinberg" <zackw@stanford.edu>]

On Fri, Jun 08, 2001 at 10:06:51AM +0300, Eli Zaretskii wrote:
> 
> One notorious problem with GNU regex is that it is quite slow for many
> simple jobs, such as matching a simple regular expression with no
> backtracking.  It seems that the main reason for this slowness is the
> fact that GNU regex supports null characters in strings.  For
> examnple, Sed 3.02 compiled with GNU regex is about 2-4 times slower
> on simple jobs than the same Sed compiled with Spencer's regex
> library.

I think the null characters are a red herring.  I looked into GNU
regex's performance in the context of GCC's fixincludes program, last
year.  On a platform that has mostly-okay headers, fixincludes spends
most of its time matching regular expressions.

The regex.c that came with GDB 4.18, which I think is the one that got
spread around widely, had a bug in its implementation of the POSIX
regcomp/regexec interface, which caused a major performance hit.  That
bug has been fixed in GNU libc for a long time.  When I replaced
fixincludes' copy of regex.c with a more recent version from glibc,
fixincludes was sped up by a factor of nine.  That same bug affects
Sed 3.02 - replace the regex.c it ships with with the one from glibc
2.2.x and I bet you'll see better performance.

There's some discussion in these messages:

http://gcc.gnu.org/ml/gcc-patches/2000-01/msg00764.html
http://gcc.gnu.org/ml/gcc-patches/2000-01/msg00765.html

The relevant fix is in there, too, if you want to pull it out and
apply it.

I did some benchmarking of fixincludes with Spencer's regexp library
as well.  IIRC, it was about the same as the fixed GNU regex.c.

-- 
zw        This is, no doubt, the rational strategy; quite possibly the
          only one that will work.  But it ignores the exigiencies of
          the tenure system and is therefore impractical.
          	-- Jerry Fodor, _The Mind Doesn't Work That Way_